A deep learning model for automated identification of age-related macular degeneration atrophy.
Authors
Affiliations (18)
Affiliations (18)
- Department of Computational Medicine, University of California Los Angeles, Los Angeles, CA, USA. [email protected].
- Department of Computer Science, University of California Los Angeles, Los Angeles, CA, USA. [email protected].
- Department of Anesthesiology and Perioperative Medicine, University of California Los Angeles, Los Angeles, CA, USA. [email protected].
- Optum AI, UnitedHealth Group, Eden Prairie, MN, USA. [email protected].
- Department of Ophthalmology, Hadassah Medical Organization and Faculty of Medicine, Hebrew University of Jerusalem, Jerusalem, Israel.
- Doheny Eye Institute, Pasadena, CA, USA.
- Department of Ophthalmology, University of California Los Angeles, Los Angeles, CA, USA.
- Department of Computational Medicine, University of California Los Angeles, Los Angeles, CA, USA.
- Stein Eye Institute, University of California Los Angeles, Los Angeles, CA, USA.
- Department of Computer Science, University of California Los Angeles, Los Angeles, CA, USA.
- Office of Health Informatics and Analytics, UCLA Health, Los Angeles, CA, USA.
- Department of Anesthesiology and Perioperative Medicine, University of California Los Angeles, Los Angeles, CA, USA.
- Retina Consultants of Texas, Retina Consultants of America, Houston, TX, USA.
- Blanton Eye Institute, Houston Methodist Hospital, Houston, TX, USA.
- Department of Computer Science, New York University, New York, NY, USA.
- Division of Precision Medicine, New York University, New York, NY, USA.
- Department of Neurosurgery, University of California Los Angeles, Los Angeles, CA, USA.
- Department of Ophthalmology, Hadassah Medical Organization and Faculty of Medicine, Hebrew University of Jerusalem, Jerusalem, Israel. [email protected].
Abstract
Age-related macular degeneration (AMD), a leading cause of visual impairment and blindness among the elderly, is projected to affect 288 million individuals globally by 2040. Advanced AMD, including complete retinal pigment epithelium and outer retinal atrophy (cRORA), pose significant challenges for diagnosis and monitoring due to the labor-intensive, costly, and variable nature of manual annotation of volumetric optical coherence tomography (OCT) scans. Automating cRORA diagnosis offers the potential to improve annotation consistency and reduce clinical burden, which could facilitate, for example, the evaluation of recently FDA-approved treatments that delay disease progression. In this study, we compiled two large independent cohorts totaling nearly 5,000 3D OCT scans, labeled them for cRORA presence, and developed a deep learning model for cRORA automated detection. The model achieved state-of-the-art performance, with a ROC AUC of 0.97 on internal validation, and demonstrated robust translatability (zero-shot learning) with a ROC AUC of 0.88 on external evaluation. Notably, it exhibited high accuracy for both non-neovascular (non-nv) and neovascular (nv) AMD subgroups (ROC AUC 0.98 and 0.93, respectively), including complex cases with exudation. This model and dataset combination could facilitate clinical research and trial analyses by providing scalable, standardized assessments across non-nv and nv AMD patient subgroups.